Re: [dpdk-dev] [PATCH v2] mempool/ring: add support for new ring sync modes

[Olivier Matz <olivier.matz@6wind.com>]

On Mon, Jul 13, 2020 at 02:46:35PM +0000, Ananyev, Konstantin wrote:
> Hi Olivier,
> 
> > Hi Konstantin,
> > 
> > On Fri, Jul 10, 2020 at 03:20:12PM +0000, Ananyev, Konstantin wrote:
> > >
> > >
> > > >
> > > > Hi Olivier,
> > > >
> > > > > Hi Konstantin,
> > > > >
> > > > > On Thu, Jul 09, 2020 at 05:55:30PM +0000, Ananyev, Konstantin wrote:
> > > > > > Hi Olivier,
> > > > > >
> > > > > > > Hi Konstantin,
> > > > > > >
> > > > > > > On Mon, Jun 29, 2020 at 05:10:24PM +0100, Konstantin Ananyev wrote:
> > > > > > > > v2:
> > > > > > > >  - update Release Notes (as per comments)
> > > > > > > >
> > > > > > > > Two new sync modes were introduced into rte_ring:
> > > > > > > > relaxed tail sync (RTS) and head/tail sync (HTS).
> > > > > > > > This change provides user with ability to select these
> > > > > > > > modes for ring based mempool via mempool ops API.
> > > > > > > >
> > > > > > > > Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> > > > > > > > Acked-by: Gage Eads <gage.eads@intel.com>
> > > > > > > > ---
> > > > > > > >  doc/guides/rel_notes/release_20_08.rst  |  6 ++
> > > > > > > >  drivers/mempool/ring/rte_mempool_ring.c | 97 ++++++++++++++++++++++---
> > > > > > > >  2 files changed, 94 insertions(+), 9 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/doc/guides/rel_notes/release_20_08.rst b/doc/guides/rel_notes/release_20_08.rst
> > > > > > > > index eaaf11c37..7bdcf3aac 100644
> > > > > > > > --- a/doc/guides/rel_notes/release_20_08.rst
> > > > > > > > +++ b/doc/guides/rel_notes/release_20_08.rst
> > > > > > > > @@ -84,6 +84,12 @@ New Features
> > > > > > > >    * Dump ``rte_flow`` memory consumption.
> > > > > > > >    * Measure packet per second forwarding.
> > > > > > > >
> > > > > > > > +* **Added support for new sync modes into mempool ring driver.**
> > > > > > > > +
> > > > > > > > +  Added ability to select new ring synchronisation modes:
> > > > > > > > +  ``relaxed tail sync (ring_mt_rts)`` and ``head/tail sync (ring_mt_hts)``
> > > > > > > > +  via mempool ops API.
> > > > > > > > +
> > > > > > > >
> > > > > > > >  Removed Items
> > > > > > > >  -------------
> > > > > > > > diff --git a/drivers/mempool/ring/rte_mempool_ring.c b/drivers/mempool/ring/rte_mempool_ring.c
> > > > > > > > index bc123fc52..15ec7dee7 100644
> > > > > > > > --- a/drivers/mempool/ring/rte_mempool_ring.c
> > > > > > > > +++ b/drivers/mempool/ring/rte_mempool_ring.c
> > > > > > > > @@ -25,6 +25,22 @@ common_ring_sp_enqueue(struct rte_mempool *mp, void * const *obj_table,
> > > > > > > >  			obj_table, n, NULL) == 0 ? -ENOBUFS : 0;
> > > > > > > >  }
> > > > > > > >
> > > > > > > > +static int
> > > > > > > > +rts_ring_mp_enqueue(struct rte_mempool *mp, void * const *obj_table,
> > > > > > > > +	unsigned int n)
> > > > > > > > +{
> > > > > > > > +	return rte_ring_mp_rts_enqueue_bulk(mp->pool_data,
> > > > > > > > +			obj_table, n, NULL) == 0 ? -ENOBUFS : 0;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > +static int
> > > > > > > > +hts_ring_mp_enqueue(struct rte_mempool *mp, void * const *obj_table,
> > > > > > > > +	unsigned int n)
> > > > > > > > +{
> > > > > > > > +	return rte_ring_mp_hts_enqueue_bulk(mp->pool_data,
> > > > > > > > +			obj_table, n, NULL) == 0 ? -ENOBUFS : 0;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > >  static int
> > > > > > > >  common_ring_mc_dequeue(struct rte_mempool *mp, void **obj_table, unsigned n)
> > > > > > > >  {
> > > > > > > > @@ -39,17 +55,30 @@ common_ring_sc_dequeue(struct rte_mempool *mp, void **obj_table, unsigned n)
> > > > > > > >  			obj_table, n, NULL) == 0 ? -ENOBUFS : 0;
> > > > > > > >  }
> > > > > > > >
> > > > > > > > +static int
> > > > > > > > +rts_ring_mc_dequeue(struct rte_mempool *mp, void **obj_table, unsigned int n)
> > > > > > > > +{
> > > > > > > > +	return rte_ring_mc_rts_dequeue_bulk(mp->pool_data,
> > > > > > > > +			obj_table, n, NULL) == 0 ? -ENOBUFS : 0;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > +static int
> > > > > > > > +hts_ring_mc_dequeue(struct rte_mempool *mp, void **obj_table, unsigned int n)
> > > > > > > > +{
> > > > > > > > +	return rte_ring_mc_hts_dequeue_bulk(mp->pool_data,
> > > > > > > > +			obj_table, n, NULL) == 0 ? -ENOBUFS : 0;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > >  static unsigned
> > > > > > > >  common_ring_get_count(const struct rte_mempool *mp)
> > > > > > > >  {
> > > > > > > >  	return rte_ring_count(mp->pool_data);
> > > > > > > >  }
> > > > > > > >
> > > > > > > > -
> > > > > > > >  static int
> > > > > > > > -common_ring_alloc(struct rte_mempool *mp)
> > > > > > > > +ring_alloc(struct rte_mempool *mp, uint32_t rg_flags)
> > > > > > > >  {
> > > > > > > > -	int rg_flags = 0, ret;
> > > > > > > > +	int ret;
> > > > > > > >  	char rg_name[RTE_RING_NAMESIZE];
> > > > > > > >  	struct rte_ring *r;
> > > > > > > >
> > > > > > > > @@ -60,12 +89,6 @@ common_ring_alloc(struct rte_mempool *mp)
> > > > > > > >  		return -rte_errno;
> > > > > > > >  	}
> > > > > > > >
> > > > > > > > -	/* ring flags */
> > > > > > > > -	if (mp->flags & MEMPOOL_F_SP_PUT)
> > > > > > > > -		rg_flags |= RING_F_SP_ENQ;
> > > > > > > > -	if (mp->flags & MEMPOOL_F_SC_GET)
> > > > > > > > -		rg_flags |= RING_F_SC_DEQ;
> > > > > > > > -
> > > > > > > >  	/*
> > > > > > > >  	 * Allocate the ring that will be used to store objects.
> > > > > > > >  	 * Ring functions will return appropriate errors if we are
> > > > > > > > @@ -82,6 +105,40 @@ common_ring_alloc(struct rte_mempool *mp)
> > > > > > > >  	return 0;
> > > > > > > >  }
> > > > > > > >
> > > > > > > > +static int
> > > > > > > > +common_ring_alloc(struct rte_mempool *mp)
> > > > > > > > +{
> > > > > > > > +	uint32_t rg_flags;
> > > > > > > > +
> > > > > > > > +	rg_flags = 0;
> > > > > > >
> > > > > > > Maybe it could go on the same line
> > > > > > >
> > > > > > > > +
> > > > > > > > +	/* ring flags */
> > > > > > >
> > > > > > > Not sure we need to keep this comment
> > > > > > >
> > > > > > > > +	if (mp->flags & MEMPOOL_F_SP_PUT)
> > > > > > > > +		rg_flags |= RING_F_SP_ENQ;
> > > > > > > > +	if (mp->flags & MEMPOOL_F_SC_GET)
> > > > > > > > +		rg_flags |= RING_F_SC_DEQ;
> > > > > > > > +
> > > > > > > > +	return ring_alloc(mp, rg_flags);
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > +static int
> > > > > > > > +rts_ring_alloc(struct rte_mempool *mp)
> > > > > > > > +{
> > > > > > > > +	if ((mp->flags & (MEMPOOL_F_SP_PUT | MEMPOOL_F_SC_GET)) != 0)
> > > > > > > > +		return -EINVAL;
> > > > > > >
> > > > > > > Why do we need this? It is a problem to allow sc/sp in this mode (even
> > > > > > > if it's not optimal)?
> > > > > >
> > > > > > These new sync modes (RTS, HTS) are for MT.
> > > > > > For SP/SC - there is simply no point to use MT sync modes.
> > > > > > I suppose there are few choices:
> > > > > > 1. Make F_SP_PUT/F_SC_GET flags silently override expected ops behaviour
> > > > > >    and create actual ring with ST sync mode for prod/cons.
> > > > > > 2. Report an error.
> > > > > > 3. Silently ignore these flags.
> > > > > >
> > > > > > As I can see for  "ring_mp_mc" ops, we doing #1,
> > > > > > while for "stack" we are doing #3.
> > > > > > For RTS/HTS I chosoe #2, as it seems cleaner to me.
> > > > > > Any thoughts from your side what preferable behaviour should be?
> > > > >
> > > > > The F_SP_PUT/F_SC_GET are only used in rte_mempool_create() to select
> > > > > the default ops among (ring_sp_sc, ring_mp_sc, ring_sp_mc,
> > > > > ring_mp_mc).
> > > >
> > > > As I understand, nothing prevents user from doing:
> > > >
> > > > mp = rte_mempool_create_empty(name, n, elt_size, cache_size,
> > > >                  sizeof(struct rte_pktmbuf_pool_private), socket_id, 0);
> > >
> > > Apologies, hit send accidently.
> > > I meant user can do:
> > >
> > > mp = rte_mempool_create_empty(..., F_SP_PUT | F_SC_GET);
> > > rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
> > >
> > > An in that case, he'll get SP/SC ring underneath.
> > 
> > It looks it's not the case. Since commit 449c49b93a6b ("mempool: support
> > handler operations"), the flags SP_PUT/SC_GET are converted into a call
> > to rte_mempool_set_ops_byname() in rte_mempool_create() only.
> > 
> > In rte_mempool_create_empty(), these flags are ignored. It is expected
> > that the user calls rte_mempool_set_ops_byname() by itself.
> 
> As I understand the code - not exactly.
> rte_mempool_create_empty() doesn't make any specific actions based on 'flags' value,
> but it does store it's value inside mp->flags.
> Later, when mempool_ops_alloc_once() is called these flags will be used by
> common_ring_alloc() and might override selected by ops ring behaviour.
> 
> > 
> > I don't think it is a good behavior:
> > 
> > 1/ The documentation of rte_mempool_create_empty() does not say that the
> >    flags are ignored, and a user can expect that F_SP_PUT | F_SC_GET
> >    sets the default ops like rte_mempool_create().
> > 
> > 2/ If rte_mempool_set_ops_byname() is not called after
> >    rte_mempool_create_empty() (and it looks it happens in dpdk's code),
> >    the default ops are the ones registered at index 0. This depends on
> >    the link order.
> > 
> > So I propose to move the following code in
> > rte_mempool_create_empty().
> > 
> > 	if ((flags & MEMPOOL_F_SP_PUT) && (flags & MEMPOOL_F_SC_GET))
> > 		ret = rte_mempool_set_ops_byname(mp, "ring_sp_sc", NULL);
> > 	else if (flags & MEMPOOL_F_SP_PUT)
> > 		ret = rte_mempool_set_ops_byname(mp, "ring_sp_mc", NULL);
> > 	else if (flags & MEMPOOL_F_SC_GET)
> > 		ret = rte_mempool_set_ops_byname(mp, "ring_mp_sc", NULL);
> > 	else
> > 		ret = rte_mempool_set_ops_byname(mp, "ring_mp_mc", NULL);
> > 
> > What do you think?
> 
> I think it will be a good thing - as in that case we'll always have
> "ring_mp_mc" selected as default one.
> As another thought, it porbably would be good to deprecate and later remove
> MEMPOOL_F_SP_PUT and MEMPOOL_F_SC_GET completely.
> These days user can select this behaviour via mempool ops and such dualism
> just makes things more error-prone and harder to maintain.
> Especially as we don't have clear policy what should be the higher priority
> for sync mode selection: mempool ops or flags. 
> 

I'll tend to agree, however it would mean deprecate rte_mempool_create()
too, because we wouldn't be able to set ops with it. Or we would have to
add a 12th (!) argument to the function, to set the ops name.

I don't like having that many arguments to this function, but it seems
it is widely used, probably because it is just one function call (vs
create_empty + set_ops + populate). So adding a "ops_name" argument is
maybe the right thing to do, given we can keep abi compat.